The objective of this project is to perform the initial development and evaluation of a computer aid to assist radiologists in their interpretation of mammograms. We will develop and evaluate an approach to computer-aided diagnosis (CAD(, in which the radiologist will be assisted by a content-based search engine that will display examples of lesions, with known pathology, that are similar to the lesion being evaluated. We will model the perceptual similarity between two lesion images as a non-linear function of those images, and use algorithms (support vector machines and artificial neural networks) to learn this function from similarity techniques that will allow the radiologist to refine the search by indicating preferences among the retrieved images, providing a capability similar to that present in text-search engines. We will focus only on the retrieval of images of microcalcification clusters (MCCs) to determine the feasibility of later developing a more-complete system capable of handling multiple lesion classes. The project will involve a thorough performance evaluation to determine the merits of continued development of the proposed approach to CAD. We will perform statistical analyses of inter-observer and intra-observer notions of image similarity, and use modern statistical resampling procedures to evaluate the generation error of our nonlinear similarity model. The specific aims of the proposed project are as follows: 1) Develop support-vector-machine and artificial-neural network methods for predicting radiologists' similarity assessments from image features extracted by computer; 2) Develop relevance-feedback techniques for refining searches based on user-assessed relevance of retrieved images; 3) Based on an MCC data set, obtain radiologists' similarity assessments, for training and testing the proposed image-retrieval system; and 4) Evaluate retrieval performance by using quantitative measures, such as precision-recall curves and generalization error, and studies of inter-observer and intra-observer variability; study diagnostic utility by measuring the fraction of retrieved images that share th same pathology as the query.